Words have a place.
Explore meaning as position — and relationships as directions.
What embeddings are these? / 数据与教学说明
Real pretrained vectors, not invented coordinates. GloVe 6B, 50 dimensions, trained on Wikipedia 2014 + Gigaword 5 (6 billion tokens). This file embeds a curated teaching subset of the original 400,000-word vocabulary. The original numeric vectors are preserved. Categories are hand-assigned display labels, not learned classes.
How the calculation works. For a word query we rank all other included words by cosine similarity. For an analogy we calculate q = v(B) − v(A) + v(C) using the original, unnormalized 50D vectors, then rank by cosine similarity to q, excluding A, B and C. Exactly three results are shown. Scores are similarities, not probabilities. Results are restricted to the embedded subset; the full vocabulary can produce different rankings.
The orange arrows show A → B and the same displacement from C to the computed target (◇). The dashed line links that target to the winning word: an approximation, not an exact equality. An analogy may fail; these static vectors reflect patterns and biases in their training text and assign only one vector to each word.
中文说明：这是真实预训练词向量的小型词表，不是手工编造的坐标。最近邻和 B − A + C 都在原始 50 维空间计算；三维 PCA 仅用于展示，不能保留所有距离与关系。结果只在本页收录词表中检索，不代表完整词库的前三名。拖动图形旋转，点击词语查看恰好三个近邻。
Source: Stanford GloVe · Pennington, Socher & Manning (2014). Data license: PDDL 1.0. Download mirror: UCL NLP. All data, rendering and calculations are embedded; these optional reference links are the only online resources.